M e d i c a l P hy s i c s a n d I n f o r m a t i c s • O r i g i n a l R e s e a r c h Joemai et al. Adaptive Iterative Dose Reduction
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 06/11/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Medical Physics and Informatics Original Research
Adaptive Iterative Dose Reduction 3D Versus Filtered Back Projection in CT: Evaluation of Image Quality Raoul M. S. Joemai1 Wouter J. H. Veldkamp1 Lucia J. M. Kroft 1 Irene Hernandez-Giron2 Jacob Geleijns1 Joemai RMS, Veldkamp WJH, Kroft LJM, Hernandez-Giron I, Geleijns J
OBJECTIVE. The purpose of this study was to evaluate image quality with filtered back projection (FBP) and adaptive iterative dose reduction 3D (AIDR 3D). MATERIALS AND METHODS. Phantom acquisitions were performed at six dose levels to assess spatial resolution, noise, and low-contrast detectability (LCD). Spatial resolution was assessed with the modulation transfer function at high and low contrast levels. Noise power spectrum and SD of attenuation were assessed. LCD was calculated with a mathematic model observer applied to phantom CT images. The subjective image quality of clinical CT scans was assessed by five radiologists. RESULTS. Compared with FBP, AIDR 3D resulted in substantial noise reduction at all frequencies with a similar shape of the noise power spectrum. Spatial resolution was similar for AIDR 3D and FBP. LCD improved with AIDR 3D, which was associated with a potential average dose reduction of 36% (range, 9–86%). The observer study showed that overall image quality improved and artifacts decreased with AIDR 3D. CONCLUSION. AIDR 3D performs better than FBP with regard to noise and LCD, resulting in better image quality, and performs similarly with respect to spatial resolution. The evaluation of image quality of clinical CT scans was consistent with the objective assessment of image quality with a phantom. The amount of dose reduction should be investigated for each clinical indication in studies with larger numbers of patients.
W
Keywords: CT, image quality, image reconstruction, medical image processing, radiation dose DOI:10.2214/AJR.12.9780 Received August 9, 2012; accepted after revision March 8, 2013. Supported by a research grant from Toshiba Medical Systems. 1 Department of Radiology, Leiden University Medical Center, C2-S, PO Box 9600, 2300RC Leiden, The Netherlands. Address correspondence to R. M. S. Joemai ([email protected]). 2
Universitat Rovira i Virgili, Física Mèdica, Reus, Spain.
AJR 2013; 201:1291–1297 0361–803X/13/2016–1291 © American Roentgen Ray Society
ith increasing recognition of the importance of radiation protection, dose reduction has become an important issue in CT system development. Modern CT systems are equipped with several dose reduction techniques. These techniques range from hardware, such as a sliding collimator to eliminate unnecessary radiation exposure due to overranging [1], to algorithms such as improved filtered back projection (FBP) [2] and iterative reconstruction [3–6]. Newer iterative reconstruction techniques are promising for effectively reducing dose. Studies have shown that with iterative techniques, image noise (measured as the SD of attenuation) decreases [3] and contrast-tonoise ratio increases compared with the values with FBP at similar dose levels [7]. Clinical observer studies have confirmed that with iterative techniques, substantial reduction in radiation dose can be achieved without loss of image quality [7]. A limitation of iterative reconstruction is that the images have a different appearance [6] resulting from a change in
shape of noise power spectrum (NPS). This change in NPS shape hampers objective comparisons of image quality by assessment of noise expressed as SD of attenuation, which is the general method for measuring noise and its relation to radiation dose. Each CT manufacturer has developed its own iterative reconstruction technique. Consequently, results may not be generalized for all iterative reconstruction techniques. Studies have been performed with ASiR (GE Healthcare) [8, 9], Veo (GE Healthcare), IRIS (Siemens Healthcare) [3, 10], Safire (Siemens Healthcare) [11, 12], iDose (Philips Healthcare) [13, 14], and AIDR (Toshiba Medical Systems) [15]. This study was performed with AIDR 3D [16–19], which is the successor to AIDR. Objective measurements in most of the previous studies have been limited to measurements of noise levels as SD [3, 4, 6]. In only a few studies [8, 13, 20, 21] did investigators support their results with NPS measurements, and some investigators [9, 13, 15, 22] added an objective evaluation of spatial resolution. In general, study results show a significant decrease in noise and
AJR:201, December 2013 1291
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 06/11/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Joemai et al. potential dose reduction. The aim of this study was to use quantitative techniques for comparison of image quality for FBP and AIDR 3D. In addition, a subjective evaluation was performed to support the quantitative findings. Materials and Methods Adaptive Iterative Dose Reduction 3D AIDR 3D (Toshiba Medical Systems) is an iterative algorithm for performing noise reduction techniques in the raw data and image domain. Processing in the raw data domain involves a statistical model, scanner model, and projection noise estimation to decrease noise [23]. Then in the image domain an iterative noise reduction technique is applied to optimize the reconstruction [24]. The AIDR 3D algorithm is implemented in the CT system and can be optionally selected. An upgrade of the reconstruction unit allows similar reconstruction time for AIDR 3D and FBP.
Phantom A phantom (Catphan 500, The Phantom Laboratory) was used for the evaluation of image quality. The homogeneous module was used to perform noise measurements and the low-contrast module was used to calculate the detectability of low-contrast objects. The low-contrast detectability (LCD) module contains three patterns of disk-shaped objects with three nominal contrast levels (1%, 0.5%, and 0.3%). In this study, objects with a nominal contrast level of 1.0% were used. From the sensitometry module, two disks consisting of different materials (low-density polyethylene and polytetrafluoroethylene [Teflon, Du-
Pont]) were used to calculate the spatial resolution with the modulation transfer function (MTF).
Acquisition and Reconstruction Protocol The phantom data were acquired with an Aquilion ONE CT scanner (version 4.74, Toshiba Medical Systems). The phantom was centered at the isocenter of the scanner with its cross section parallel to the axial plane. Acquisitions of the phantom were performed at 120-kV tube voltage, 64 × 0.5 mm beam collimation, scan FOV 320 mm, 0.5-second rotation time, and a pitch of 0.83. Acquisitions were performed at six tube current levels: 500, 300, 150, 80, 40, and 20 mA. Phantom images were reconstructed at 0.5-mm slice thickness. The reconstruction interval was 0.5 mm, and the reconstruction FOV was 240 mm. The images were reconstructed with two kernels, one soft body reconstruction kernel (FC13) and one slightly sharper body reconstruction kernel (FC14). All volumes were reconstructed by conventional FBP and AIDR 3D. Ten clinical studies were selected, and corresponding acquisition and reconstruction parameters were determined by the clinical acquisition protocol (Table 1). This included automatic tube current modulation for acquisition protocols of the thorax and abdomen. The settings for automatic exposure control were different and depended on the clinical indication.
Objective Image Quality Evaluation Analysis of the images was performed with algorithms developed in Matlab (Matlab R2011a, MathWorks).
Noise—Noise was characterized through calculation of the NPS [25, 26] from CT images of the homogeneous module of the phantom. The NPS was calculated by extracting a centered 128 × 128 matrix in 50 slices. Next, the mean pixel value of the extracted matrix was subtracted from the matrix to avoid an offset in the Fourier transform. The extracted matrix was then zero-padded to a 512 × 512 image to achieve increased resolution in the Fourier domain, which results in a smoother NPS. The Fourier transform of each extended matrix and the square of the magnitude of the Fourier transform were calculated. In addition, the usual NPS normalization was performed with the ratio between pixel spacing and image size products in the horizontal and vertical directions (512 × 512). To improve accuracy, the result was averaged over all 50 slices. This approach resulted in a 2D NPS. To present these results as a 1D NPS with reduced statistical variation, the NPS was radially averaged. The same region of interest used for the NPS measurements was used to calculate the SD of attenuation. Spatial resolution—The algorithm developed was based on the method described in 2012 by Richard et al. [27]. The MTF was calculated with two disk-shaped objects: one with a low contrast level (low-density polyethylene) and one with a high contrast level (Teflon). The edge spread function was determined from the edges of these disk-shaped objects. The line spread function was calculated by the derivative of the edge spread function. Fourier transformation of the line spread function yielded the MTF. The spatial frequency at which the MTF was 50% was used as a measure to compare the spatial resolution between FBP and AIDR 3D.
TABLE 1: Acquisition Parameters for Clinical Studies Used in Subjective Evaluation
Parameter
Unenhanced Head
Contrast-Enhanced Cardiac Coronary CT Angiography
Unenhanced Chest Contrast-Enhanced Contrast-Enhanced Unenhanced Unenhanced Metastasis Chest and Abdomen Abdominal Abscess Knee Shoulder
Acquisition protocol Tube voltage (kV) Tube current (mA) Beam collimation (mm)
120
120
250
250
32 × 0.5 32 × 0.5
100
100
120
120
120
120
120
135
550
520
AEC
AEC
AEC
AEC
80
150
280 × 0.5
280 × 0.5
100 × 0.5
160 × 0.5
160 × 0.5
160 × 0.5
64 × 0.5
32 × 0.5
1.39
0.869
0.869
0.869
1.28
0.656
Pitch
0.656
0.656
0
0
Rotation time (s)
0.75
0.75
0.35
0.35
0.5
0.5
0.5
0.5
0.5
0.5
Scan FOV (mm)
320
320
320
320
400
400
400
400
500
500
Reconstruction FOV (mm)
248.75
215
181.25
215
300
378.1
400
332.8
219
500
Kernel
FC26
FC26
FC03
FC03
FC12
FC12
FC12
FC12
FC04 and FC30
FC04 and FC30
Slice thickness (mm)
0.5
0.5
0.5
0.5
1
1
1
1
0.5
1
Slice interval (mm)
0.5
0.5
0.3
0.3
1
1
1
1
0.4
1
1194.4
172.9
159.9
76.2
781.8
991.1
405.5
110.9
331.8
Reconstruction protocol
Radiation dose Dose-length product (mGy ⋅ cm) 1163.2
Note—AEC = automatic exposure control.
1292
AJR:201, December 2013
Low-contrast detectability—The low-contrast sensitivity of CT scanners is routinely assessed by subjective scoring of low-contrast objects within phantom CT images. The theoretic difference in attenuation between the low-contrast objects (disks) and background in the phantom was 10, 5, and 3 HU for the 1%, 0.5%, and 0.3% contrast series. Therefore, noise, resolution, and contrast were significant factors in these measurements. Rating of low-contrast visibility by human observers can be biased because low-contrast objects are arranged in known fixed patterns. As an alternative to analysis by human observers, a nonprewhitening matched filter with eye filter model observer was applied in a two-alternative forced-choice experiment [28]. The nonprewhitening matched filter with eye filter is a mathematic model that has been found similar to human observers for detection tasks in the presence of low-pass noise [29]. Its strategy consists of correlating the image with the shape of the expected signal profile (of background and signal) filtered by the visual-response function. In a detection task, the model reaches a decision by comparing test statistics T1 (test for background) and T2 (test for signal). The test statistics are ob1.2
tained by cross correlation between the expected signal and the image [30]. An average slice was created with all the slices available for each individual dataset. Template locations (defined according to manufacturer specifications) were shifted in a 3 × 3 pixel region centered at the initial template location to find the maximum value (average pixel value within the template). The location that gave the maximum value was assigned as the new (optimized) template location. The amplitude of the signals in the template was fixed for each contrast series and based on specifications. The signal is blurred by the point spread function measured with the appropriate phantom. The eye filter, E, used in the model was E(f) = fe–bf with b chosen such that E(f) peaked at four cycles per degree. The eye filter was radially symmetric, f is spatial frequency, and e is Euler’s number. In the experiments, a fixed viewing distance of 500 mm from the monitor was assumed. From the distribution of test statistics, one can compute a discrimination index. Discrimination indexes (d′) resulting from the automated analysis were transformed into proportion correct responses varying between 0.5 (chance) and 1.0 (100% detectability)
FBP FC13
Normalized NPS
Normalized NPS
Subjective Image Quality Evaluation Subjective image quality was evaluated for 10 retrospectively selected clinical datasets. All patients
FBP FC13
AIDR 3D FC14
1.0
0.8 0.6 0.4
0.8 0.6 0.4 0.2
0.2 0.0
in a two-alternative forced-choice experiment. Objects were considered visible when the proportion correct was 75% or greater. By use of a psychometric fit to the proportion correct data points, a continuous function was obtained describing proportion correct (0.5–1.0) versus diameter (zero to infinity). The psychometric fits were performed by applying a Boltzmann fit. In our study we used 41 uncorrelated slices to obtain the discrimination index for each object. This was done 20 times for each individual datum based on a bootstrap technique. In this way the SD of just-visible object diameters could be determined as a measure of accuracy for each individual dataset. Bootstrapping is a method for assigning measures of accuracy to sample estimates. It was implemented as follows: for each individual dataset of 41 slices, 20 new datasets were constructed (equal size to the original dataset), each of which was obtained by random sampling with replacement from the original dataset (some slices present more than once in a new dataset).
1.2
AIDR 3D FC13
1.0
0
2
4
6 8 10 Resolution (line pairs cm)
12
0.0
14
0
2
4
6 8 10 Resolution (line pairs cm)
12
14
A
B
Fig. 1—Graphs show noise power spectrum (NPS) of filtered back projection (FBP) and adaptive iterative dose reduction 3D (AIDR 3D) at tube charge 75 mAs at same reconstruction kernel (FC13) (A) and at different reconstruction kernels (FC13 and FC14) (B).
0.018
0.018
FBP FC13 FBP FC14
0.016
SD of Attenuation–2 (HU)
SD of Attenuation–2 (HU)
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 06/11/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Adaptive Iterative Dose Reduction
0.014 0.012 0.010 0.008 0.006 0.004 0.002 0
0
50
100 150 200 Tube Charge (mAs)
250
300
AIDR 3D FC13 AIDR 3D FC14
0.016 0.014 0.012 0.010 0.008 0.006 0.004 0.002 0
0
50
100 200 150 Tube Charge (mAs)
250
300
A
B
Fig. 2—Graphs show inverse square of SD versus tube charge for filtered back projection (FBP) (A) and adaptive iterative dose reduction 3D (AIDR 3D) (B) at different reconstruction kernels (FC13 and FC14).
AJR:201, December 2013 1293
Joemai et al. AIDR3D
Teflon Kernel FC13
0.8
0.8 f50 (lp/mm)
f50 (lp/mm)
FBP
1.0
0.6 0.4 0.2 0.0
0
50
100 150 200 250 Tube Charge (mAs)
0.4
FBP
0.0
300 AIDR3D
0.8
0.8
0.6 0.4
0
50
100 150 200 250 Tube Charge (mAs)
Teflon Kernel FC14
1.0
f50 (lp/mm)
f50 (lp/mm)
lished by the referring physician was used. Thus the image quality score was determined for the assessment of specific clinical indications (e.g., possible metastasis, bleeding, abscess, coronary stenosis, and fractures). According to the indication, a diagnostic quality score was determined by the observers. The observers were notified that artifacts caused by motion, such as respiratory motion or cardiac motion, were excluded from the scoring. The following technical artifacts were assessed: helical or windmill artifacts, streak artifacts, and beam-hardening artifacts. Observers were instructed to express every perceived difference in image quality between FBP and AIDR 3D in five grades to record the results on a score chart. Adjustments in window settings and zooming were allowed in image evaluation. For evaluation of the subjective image quality of the clinical scans for FBP and AIDR 3D, for each CT scan the mean observer score was calculated. These mean observer scores were compared on a pairwise basis by Wilcoxon signed rank test. Variations in statistical analysis due to the different observers and clinical datasets were minimized by use of the mean observer score.
AIDR3D
0.2
LDPE Kernel FC14
FBP
300
AIDR3D
0.6 0.4 0.2
0.2 0.0
FBP
0.6
1.0
0
50
100 150 200 250 Tube Charge (mAs)
300
0.0
0
50
100 150 200 250 Tube Charge (mAs)
300
Fig. 3—Graphs of spatial resolution expressed as spatial frequency in which modulation transfer function is 50% (f 50) for filtered back projection (FBP) and adaptive iterative dose reduction 3D (AIDR 3D) show results for two contrast levels (low-density polyethylene [LDPE] and polytetrafluoroethylene [Teflon, DuPont]) and two reconstruction kernels (FC13 and FC14).
had undergone a CT examination for clinical indications according to the clinical acquisition protocol. Datasets were retrospectively selected from the most recent examinations only on the basis of the scanned region. Two studies were consecutively selected for each region: head, chest, cardiac, abdomen, and extremities. Our institutional review board does not require approval for retrospective analysis of data with patient-identifying information omitted, as was the case in this study. Five radiologists experienced in clinical CT evaluated the images. Image evaluation was performed at a dedicated medical image postprocessing workstation (Vitrea fX version 6.1, Vital Images). For
8 7
each clinical dataset, conventional FBP and AIDR 3D volumes were loaded in the workstation and presented simultaneously. The observers were blinded to reconstruction technique. The datasets were displayed next to each other at random. The radiologists assessed image quality by scrolling through the transaxial slices of two volumes (FBP and AIDR 3D) simultaneously. They graded image quality using a qualitative five-point ranking of diagnostic quality (1, excellent; 2, good; 3, moderate; 4, limited; and 5, nondiagnostic) and artifacts (1, none; 2, minor; 3, moderate; 4, severe; and 5, unacceptable). To obtain a score of diagnostic image quality, the actual indication for each CT examination estab-
FBP FC13 AIDR3D FC13
Just Visible Diameter (mm)
Just Visible Diameter (mm)
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 06/11/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
LDPE Kernel FC13
1.0
6 5 4 3 2 1 0
250
150
75 40 Tube Charge (mAs)
20
10
8 7
Results Noise NPS graphs were normalized to the maximum. These graphs showed that at equal reconstruction kernels, a small change in noise can be expected (Fig. 1A). However, NPS measurements also showed that differences in NPS shape can be reduced by use of a slightly sharper reconstruction kernel for AIDR 3D (Fig. 1B). Subjectively, noise appeared with a lower magnitude with AIDR 3D than with FBP. With both reconstruction techniques a slight increase in noise was seen when a sharper reconstruction kernel was used. It is expected that the dose is proportional to SD −2 [31]. This linear relation was tested
FBP FC14 AIDR3D FC14
6 5 4 3 2 1 0
250
150
75 40 Tube Charge (mAs)
20
A
10
B
Fig. 4—Diameters of just-visible low-contrast object for six dose levels with filtered back projection (FBP) and adaptive iterative dose reduction 3D (AIDR 3D). A, Graph shows results with reconstruction kernel FC13. B, Graph shows results for reconstruction kernel FC14.
1294
AJR:201, December 2013
for FBP and AIDR 3D (Fig. 2). FBP had the expected relation for both reconstruction kernels. However, AIDR 3D exhibited a different behavior. Data at tube charges (i.e., tube current–time product) of 75–10 mAs showed a relatively smaller increase in noise at each step in lower-dose images compared with images obtained at high tube charges (250– 75 mAs). At all dose levels, lower noise levels were found for the slightly smoother kernel (FC13) than for the sharper FC14 kernel. Spatial Resolution Spatial resolution in which MTF is 50% is shown for two contrast levels, two reconstruction techniques, and two reconstruction kernels in Figure 3. Quantitative analysis of MTF measurements showed similar performance for AIDR 3D compared with FBP. At low dose levels (< 50 mAs), however, a small decrease in spatial resolution was found with AIDR 3D. The sharper kernel (FC14) showed a slightly better resolution than the FC13 kernel. The spatial resolution was not dependent on contrast level for AIDR 3D or for FBP. Low-Contrast Detectability Figure 4 shows the diameter of the just-visible low-contrast objects (alpha), as a function of the dose level for two reconstruction kernels. As could be expected, the LCD improved rapidly at higher dose levels, ranging for AIDR 3D from detectability of objects as
Fig. 5—Graph shows image quality score averaged for individual observers. Diagnostic quality and artifacts were rated on 5-point scale (lower is better). AIDR 3D = adaptive iterative dose reduction 3D, FBP = filtered back projection.
Mean Score (Lower Is Better)
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 06/11/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Adaptive Iterative Dose Reduction 4.0 3.5 3.0 2.5 2.0 1.5 1.0
FBP AIDR 3D Diagnostic Quality
Observer 1
Observer 2
small as 1.37 mm in diameter at the highest dose level (250 mAs, FC13) to slightly greater than 4 mm in diameter at the lowest level (10 mAs, FC13). At all dose levels there was consistent improvement in the detectability of low-contrast disks with AIDR 3D compared with FBP. This means that with AIDR 3D smaller low-contrast objects could be detected at the same dose level. LCD for the FC13 reconstruction kernel was slightly better because of a lower noise level compared with the FC14 reconstruction kernel. The relative differences between FBP and AIDR 3D differed for each dose level. The best relative improvement with AIDR 3D was found at a tube charge of 10 mAs, at which 32%-smaller-diameter low-contrast disks could be detected (Fig. 4A and 4B). The best absolute improvements were found at the lowest dose
Fig. 6—Clinical examples of data used in subjective evaluation. Left, filtered back projection; right, adaptive iterative dose reduction 3D. Top, abdomen of 47-year-old man with suspected metastasis (window width, 500 HU; level, 100 HU). Bottom, shoulders of 27-year-old woman with suspected shoulder fracture (window width, 1500 HU; level, 200 HU). Slice thickness is 0.5 mm. Substantial reduction in noise and streak artifacts is evident with AIDR 3D.
Observer 3
FBP
AIDR 3D Artifacts
Observer 4
Observer 5
levels for the FC14 reconstruction kernel, with which 2.1-mm-smaller disks could be detected with AIDR 3D (Fig. 4B). Subjective Image Quality The mean scores per reconstruction technique are shown for each observer in Figure 5. The evaluation showed the same trend for all observers with an observer-dependent bias. Individual observer results showed the best diagnostic image quality and fewest artifacts for AIDR 3D. The average score across all cases and observers for diagnostic quality improved significantly from 2.4 ± 0.6 for FBP to 1.5 ± 0.3 for AIDR 3D (p < 0.0003). The amount of artifacts also decreases significantly with AIDR 3D (2.9 ± 0.7 for FBP, 1.7 ± 0.4 for AIDR 3D; p < 0.0003). An example of the performance of AIDR 3D compared with FBP is depicted in Figure 6, which shows that a marked decrease in artifacts can be realized with AIDR 3D. Discussion Objective and subjective evaluations showed consistently that image quality improves with AIDR 3D compared with FBP. Measurements revealed that LCD improves a maximum of 32% with a substantial reduction of noise with AIDR 3D compared with FBP. Subjective grading of clinical images of patients confirmed improved image quality with reduction of artifacts and potential for dose reduction of AIDR 3D. Further clinical studies should be performed to evaluate the potential dose reduction for each particular clinical indication. There is a need for methods of accurate assessment of the diagnostic performance of new imaging technologies, particularly in CT. Our approach to the assessment of spatial resolution (MTF), noise (NPS), and LCD with a human observer model is in line with the findings of subjective grading of clinical images by radiologists. Further research and develop-
AJR:201, December 2013 1295
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 06/11/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Joemai et al. ment of such methods could be of relevance for the assessment of coming advances in imaging technologies. The human observer model seems particularly promising for achieving this objective. In addition to noise measurements as SD, it is also essential to analyze the magnitude of the noise as a function of frequency on CT images. SD measurements are commonly used but have the limitation that they do not provide information on the noise appearance. As an alternative, NPS measurements are a good tool for this task. A change in the shape of the NPS indicates that the noise appearance is different. In this study the change in the NPS shape was modest, but there was substantial reduction in the magnitude of the noise. Software for automated assessment of LCD of images of a phantom was used in this study. A previous validation study [28] showed that the software model observer has high agreement with human observers. The software, as an alternative to human observers, supports objective analysis of the LCD. Figure 4 shows that diameters corresponding to the just-visible object are larger for FBP compared with AIDR 3D for 1% contrast. From these diameters the dose reduction potential of AIRD 3D can be derived while LCD is maintained, because by interpolation, the dose level that would have been required with AIDR 3D to achieve the same alpha as for FBP can be calculated [28]. For acquisitions between 20 and 250 mAs, the calculated potential dose reduction was on average 36% (range, 9–86%). Spatial resolution measurements showed no deterioration at moderate and high dose levels, but there was improvement in noise. To achieve improvement for both spatial resolution and noise, a sharper kernel can be used in combination with AIDR 3D. In this study we used a smooth (FC13) and a slightly sharper kernel (FC14). Improvement in overall image quality is a general finding in studies comparing FBP with iterative reconstruction methods, but one of the major effects of AIDR 3D was the reduction of streak artifacts on clinical CT scans, which was achieved by processing in the raw data domain. Our study had several limitations with regard to the observer study. First, the observer study was based on subjective reading, because for practical reasons a lesion detection study, which would have been more accurate, could not be performed. Second, the subjective evaluation of clinical CT images was performed side by side for the FBP and AIDR 3D images. This method introduced the dis-
1296
advantage that observers might have been able to identify the AIDR 3D image, which might have created bias. However, the results in the subjective grading were in agreement with the objective results. Side-by-side comparison was performed because our main interest was to study the difference between the two reconstruction methods, which was considered easier to achieve in direct comparison of images. The amount of dose reduction should be investigated for each clinical indication in studies with larger numbers of patients. Finally, CT examinations of the brain and abdomen are often performed with a slice thickness of 5 mm, whereas in this study 0.5- or 1.0-mm slices were used. Because this study evaluated differences between two reconstruction techniques, thin slices were used to enhance differences between the two techniques considered. Conclusion This study showed that AIDR 3D images, compared with FBP images, have lower noise, better LCD, significantly better diagnostic image quality, and significantly fewer artifacts at a similar spatial resolution and allow a potential dose reduction of 36% (range, 9–86%) to maintain the same LCD. Acknowledgments We thank A. J. van der Molen, E. L. van Persijn-van Meerten, A. Navas Canete, and M. C. Kruit for contributions to the subjective evaluation of clinical images. References 1. Deak PD, Langner O, Lell M, et al. Effects of adaptive section collimation on patient radiation dose in multisection spiral CT. Radiology 2009; 252:140–147 2. Noo F, Hoppe S, Dennerlein F, et al. A new scheme for view-dependent data differentiation in fan-beam and cone-beam computed tomography. Phys Med Biol 2007; 52:5393–5414 3. Bittencourt MS, Schmidt B, Seltmann M, et al. Iterative reconstruction in image space (IRIS) in cardiac computed tomography: initial experience. Int J Cardiovasc Imaging 2011; 27:1081–1087 4. Flicek KT, Hara AK, Silva AC, et al. Reducing the radiation dose for CT colonography using adaptive statistical iterative reconstruction: a pilot study. AJR 2010; 195:126–131 5. Gosling O, Loader R, Venables P, et al. A comparison of radiation doses between state-of-the-art multislice CT coronary angiography with iterative reconstruction, multislice CT coronary angiography with standard filtered back-projection and invasive diagnostic coronary angiography. Heart 2010; 96:922–926
6. Leipsic J, Labounty TM, Heilbron B, et al. Adaptive statistical iterative reconstruction: assessment of image noise and image quality in coronary CT angiography. AJR 2010; 195:649–654 7. Cornfeld D, Israel G, Detroy E, et al. Impact of adaptive statistical iterative reconstruction (ASIR) on radiation dose and image quality in aortic dissection studies: a qualitative and quantitative analysis. AJR 2011; 196:[web]W336–W340 8. Marin D, Nelson RC, Schindera ST, et al. Lowtube-voltage, high-tube-current multidetector abdominal CT: improved image quality and decreased radiation dose with adaptive statistical iterative reconstruction algorithm–initial clinical experience. Radiology 2010; 254:145–153 9. Singh S, Kalra MK, Hsieh J, et al. Abdominal CT: comparison of adaptive statistical iterative and filtered back projection reconstruction techniques. Radiology 2010; 257:373–383 10. Schindera ST, Diedrichsen L, Muller HC, et al. Iterative reconstruction algorithm for abdominal multidetector CT at different tube voltages: assessment of diagnostic accuracy, image quality, and radiation dose in a phantom study. Radiology 2011; 260:454–462 11. Winklehner A, Karlo C, Puippe G, et al. Raw data-based iterative reconstruction in body CTA: evaluation of radiation dose saving potential. Eur Radiol 2011; 21:2521–2526 12. Moscariello A, Takx RA, Schoepf UJ, et al. Coronary CT angiography: image quality, diagnostic accuracy, and potential for radiation dose reduction using a novel iterative image reconstruction technique—comparison with traditional filtered back projection. Eur Radiol 2011; 21:2130–2138 13. Funama Y, Taguchi K, Utsunomiya D, et al. Combination of a low-tube-voltage technique with hybrid iterative reconstruction (iDose) algorithm at coronary computed tomographic angiography. J Comput Assist Tomogr 2011; 35:480–485 14. Habets J, Symersky P, de Mol BA, Mali WP, Leiner T, Budde RP. A novel iterative reconstruction algorithm allows reduced dose multidetector-row CT imaging of mechanical prosthetic heart valves. Int J Cardiovasc Imaging 2012; 28:1567–1575 15. Gervaise A, Osemont B, Lecocq S, et al. CT image quality improvement using adaptive iterative dose reduction with wide-volume acquisition on 320-detector CT. Eur Radiol 2012; 22:295–301 16. Yamada Y, Jinzaki M, Hosokawa T, et al. Dose reduction in chest CT: comparison of the adaptive iterative dose reduction 3D, adaptive iterative dose reduction, and filtered back projection reconstruction techniques. Eur J Radiol 2012; 81:4185–4195 17. Yoo RE, Park EA, Lee W, et al. Image quality of adaptive iterative dose reduction 3D of coronary CT angiography of 640-slice CT: comparison with filtered back-projection. Int J Cardiovasc Im-
AJR:201, December 2013
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 06/11/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Adaptive Iterative Dose Reduction aging 2013; 29:669–676 18. Ohno Y, Takenaka D, Kanda T, et al. Adaptive iterative dose reduction using 3D processing for reduced- and low-dose pulmonary CT: comparison with standard-dose CT for image noise reduction and radiological findings. AJR 2012; 199[web]:W477–W485 19. Juri H, Matsuki M, Itou Y, et al. Initial experience with adaptive iterative dose reduction 3D to reduce radiation dose in computed tomographic urography. J Comput Assist Tomogr 2013; 37:52–57 20. Ghetti C, Ortenzia O, Serreli G. CT iterative reconstruction in image space: a phantom study. Phys Med 2012; 28:161–165 21. Sagara Y, Hara AK, Pavlicek W, et al. Abdominal CT: comparison of low-dose CT with adaptive statistical iterative reconstruction and routinedose CT with filtered back projection in 53 patients. AJR 2010; 195:713–719 22. Prakash P, Kalra MK, Ackman JB, et al. Diffuse lung disease: CT of the chest with adaptive statis-
tical iterative reconstruction technique. Radiology 2010; 256:261–269 23. Yang Z, Zamyatin AA, Akino N. Effective datadomain noise and streak reduction for x-ray CT. In: 11th international meeting on fully three-dimensional image reconstruction in radiology and nuclear medicine 2011:290–296. Fully3D website. http://www.fully3d.org/2011/Fully3D2011Proceedings.pdf. Accessed August 5, 2013 24. Yang Z, Silver MD, Noshi Y. Adaptive weighted anisotropic diffusion for computed tomography denoising. In: 11th international meeting on fully three-dimensional image reconstruction in radiology and nuclear medicine 2011:210–214. Fully3D website. http://www.fully3d.org/2011/Fully3D2011Proceedings.pdf. Accessed August 5, 2013 25. Boedeker KL, Cooper VN, McNitt-Gray MF. Application of the noise power spectrum in modern diagnostic MDCT. Part 1. Measurement of noise power spectra and noise equivalent quanta. Phys Med Biol 2007; 52:4027–4046
26. Joemai RM, Geleijns J, Veldkamp WJ. Development and validation of a low dose simulator for computed tomography. Eur Radiol 2010; 20:958–966 27. Richard S, Husarik DB, Yadava G, et al. Towards task-based assessment of CT performance: system and object MTF across different reconstruction algorithms. Med Phys 2012; 39:4115–4122 28. Hernandez-Giron I, Geleijns J, Calzado A, et al. Automated assessment of low contrast sensitivity for CT systems using a model observer. Med Phys 2011; 38:S25–S35 29. Beutel J, Kundel HL, Van Metter RL. Handbook of medical imaging. Volume 1. Physics and psychophysics. Bellingham, WA: SPIE, 2000 30. Reiser I, Nishikawa RM. Identification of simulated microcalcifications in white noise and mammographic backgrounds. Med Phys 2006; 33:2905–2911 31. Primak AN, McCollough CH, Bruesewitz MR, et al. Relationship between noise, dose, and pitch in cardiac multi-detector row CT. RadioGraphics 2006; 26:1785–1794
AJR:201, December 2013 1297
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 06/11/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Medical Physics and Informatics Original Research
Adaptive Iterative Dose Reduction 3D Versus Filtered Back Projection in CT: Evaluation of Image Quality Raoul M. S. Joemai1 Wouter J. H. Veldkamp1 Lucia J. M. Kroft 1 Irene Hernandez-Giron2 Jacob Geleijns1 Joemai RMS, Veldkamp WJH, Kroft LJM, Hernandez-Giron I, Geleijns J
OBJECTIVE. The purpose of this study was to evaluate image quality with filtered back projection (FBP) and adaptive iterative dose reduction 3D (AIDR 3D). MATERIALS AND METHODS. Phantom acquisitions were performed at six dose levels to assess spatial resolution, noise, and low-contrast detectability (LCD). Spatial resolution was assessed with the modulation transfer function at high and low contrast levels. Noise power spectrum and SD of attenuation were assessed. LCD was calculated with a mathematic model observer applied to phantom CT images. The subjective image quality of clinical CT scans was assessed by five radiologists. RESULTS. Compared with FBP, AIDR 3D resulted in substantial noise reduction at all frequencies with a similar shape of the noise power spectrum. Spatial resolution was similar for AIDR 3D and FBP. LCD improved with AIDR 3D, which was associated with a potential average dose reduction of 36% (range, 9–86%). The observer study showed that overall image quality improved and artifacts decreased with AIDR 3D. CONCLUSION. AIDR 3D performs better than FBP with regard to noise and LCD, resulting in better image quality, and performs similarly with respect to spatial resolution. The evaluation of image quality of clinical CT scans was consistent with the objective assessment of image quality with a phantom. The amount of dose reduction should be investigated for each clinical indication in studies with larger numbers of patients.
W
Keywords: CT, image quality, image reconstruction, medical image processing, radiation dose DOI:10.2214/AJR.12.9780 Received August 9, 2012; accepted after revision March 8, 2013. Supported by a research grant from Toshiba Medical Systems. 1 Department of Radiology, Leiden University Medical Center, C2-S, PO Box 9600, 2300RC Leiden, The Netherlands. Address correspondence to R. M. S. Joemai ([email protected]). 2
Universitat Rovira i Virgili, Física Mèdica, Reus, Spain.
AJR 2013; 201:1291–1297 0361–803X/13/2016–1291 © American Roentgen Ray Society
ith increasing recognition of the importance of radiation protection, dose reduction has become an important issue in CT system development. Modern CT systems are equipped with several dose reduction techniques. These techniques range from hardware, such as a sliding collimator to eliminate unnecessary radiation exposure due to overranging [1], to algorithms such as improved filtered back projection (FBP) [2] and iterative reconstruction [3–6]. Newer iterative reconstruction techniques are promising for effectively reducing dose. Studies have shown that with iterative techniques, image noise (measured as the SD of attenuation) decreases [3] and contrast-tonoise ratio increases compared with the values with FBP at similar dose levels [7]. Clinical observer studies have confirmed that with iterative techniques, substantial reduction in radiation dose can be achieved without loss of image quality [7]. A limitation of iterative reconstruction is that the images have a different appearance [6] resulting from a change in
shape of noise power spectrum (NPS). This change in NPS shape hampers objective comparisons of image quality by assessment of noise expressed as SD of attenuation, which is the general method for measuring noise and its relation to radiation dose. Each CT manufacturer has developed its own iterative reconstruction technique. Consequently, results may not be generalized for all iterative reconstruction techniques. Studies have been performed with ASiR (GE Healthcare) [8, 9], Veo (GE Healthcare), IRIS (Siemens Healthcare) [3, 10], Safire (Siemens Healthcare) [11, 12], iDose (Philips Healthcare) [13, 14], and AIDR (Toshiba Medical Systems) [15]. This study was performed with AIDR 3D [16–19], which is the successor to AIDR. Objective measurements in most of the previous studies have been limited to measurements of noise levels as SD [3, 4, 6]. In only a few studies [8, 13, 20, 21] did investigators support their results with NPS measurements, and some investigators [9, 13, 15, 22] added an objective evaluation of spatial resolution. In general, study results show a significant decrease in noise and
AJR:201, December 2013 1291
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 06/11/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Joemai et al. potential dose reduction. The aim of this study was to use quantitative techniques for comparison of image quality for FBP and AIDR 3D. In addition, a subjective evaluation was performed to support the quantitative findings. Materials and Methods Adaptive Iterative Dose Reduction 3D AIDR 3D (Toshiba Medical Systems) is an iterative algorithm for performing noise reduction techniques in the raw data and image domain. Processing in the raw data domain involves a statistical model, scanner model, and projection noise estimation to decrease noise [23]. Then in the image domain an iterative noise reduction technique is applied to optimize the reconstruction [24]. The AIDR 3D algorithm is implemented in the CT system and can be optionally selected. An upgrade of the reconstruction unit allows similar reconstruction time for AIDR 3D and FBP.
Phantom A phantom (Catphan 500, The Phantom Laboratory) was used for the evaluation of image quality. The homogeneous module was used to perform noise measurements and the low-contrast module was used to calculate the detectability of low-contrast objects. The low-contrast detectability (LCD) module contains three patterns of disk-shaped objects with three nominal contrast levels (1%, 0.5%, and 0.3%). In this study, objects with a nominal contrast level of 1.0% were used. From the sensitometry module, two disks consisting of different materials (low-density polyethylene and polytetrafluoroethylene [Teflon, Du-
Pont]) were used to calculate the spatial resolution with the modulation transfer function (MTF).
Acquisition and Reconstruction Protocol The phantom data were acquired with an Aquilion ONE CT scanner (version 4.74, Toshiba Medical Systems). The phantom was centered at the isocenter of the scanner with its cross section parallel to the axial plane. Acquisitions of the phantom were performed at 120-kV tube voltage, 64 × 0.5 mm beam collimation, scan FOV 320 mm, 0.5-second rotation time, and a pitch of 0.83. Acquisitions were performed at six tube current levels: 500, 300, 150, 80, 40, and 20 mA. Phantom images were reconstructed at 0.5-mm slice thickness. The reconstruction interval was 0.5 mm, and the reconstruction FOV was 240 mm. The images were reconstructed with two kernels, one soft body reconstruction kernel (FC13) and one slightly sharper body reconstruction kernel (FC14). All volumes were reconstructed by conventional FBP and AIDR 3D. Ten clinical studies were selected, and corresponding acquisition and reconstruction parameters were determined by the clinical acquisition protocol (Table 1). This included automatic tube current modulation for acquisition protocols of the thorax and abdomen. The settings for automatic exposure control were different and depended on the clinical indication.
Objective Image Quality Evaluation Analysis of the images was performed with algorithms developed in Matlab (Matlab R2011a, MathWorks).
Noise—Noise was characterized through calculation of the NPS [25, 26] from CT images of the homogeneous module of the phantom. The NPS was calculated by extracting a centered 128 × 128 matrix in 50 slices. Next, the mean pixel value of the extracted matrix was subtracted from the matrix to avoid an offset in the Fourier transform. The extracted matrix was then zero-padded to a 512 × 512 image to achieve increased resolution in the Fourier domain, which results in a smoother NPS. The Fourier transform of each extended matrix and the square of the magnitude of the Fourier transform were calculated. In addition, the usual NPS normalization was performed with the ratio between pixel spacing and image size products in the horizontal and vertical directions (512 × 512). To improve accuracy, the result was averaged over all 50 slices. This approach resulted in a 2D NPS. To present these results as a 1D NPS with reduced statistical variation, the NPS was radially averaged. The same region of interest used for the NPS measurements was used to calculate the SD of attenuation. Spatial resolution—The algorithm developed was based on the method described in 2012 by Richard et al. [27]. The MTF was calculated with two disk-shaped objects: one with a low contrast level (low-density polyethylene) and one with a high contrast level (Teflon). The edge spread function was determined from the edges of these disk-shaped objects. The line spread function was calculated by the derivative of the edge spread function. Fourier transformation of the line spread function yielded the MTF. The spatial frequency at which the MTF was 50% was used as a measure to compare the spatial resolution between FBP and AIDR 3D.
TABLE 1: Acquisition Parameters for Clinical Studies Used in Subjective Evaluation
Parameter
Unenhanced Head
Contrast-Enhanced Cardiac Coronary CT Angiography
Unenhanced Chest Contrast-Enhanced Contrast-Enhanced Unenhanced Unenhanced Metastasis Chest and Abdomen Abdominal Abscess Knee Shoulder
Acquisition protocol Tube voltage (kV) Tube current (mA) Beam collimation (mm)
120
120
250
250
32 × 0.5 32 × 0.5
100
100
120
120
120
120
120
135
550
520
AEC
AEC
AEC
AEC
80
150
280 × 0.5
280 × 0.5
100 × 0.5
160 × 0.5
160 × 0.5
160 × 0.5
64 × 0.5
32 × 0.5
1.39
0.869
0.869
0.869
1.28
0.656
Pitch
0.656
0.656
0
0
Rotation time (s)
0.75
0.75
0.35
0.35
0.5
0.5
0.5
0.5
0.5
0.5
Scan FOV (mm)
320
320
320
320
400
400
400
400
500
500
Reconstruction FOV (mm)
248.75
215
181.25
215
300
378.1
400
332.8
219
500
Kernel
FC26
FC26
FC03
FC03
FC12
FC12
FC12
FC12
FC04 and FC30
FC04 and FC30
Slice thickness (mm)
0.5
0.5
0.5
0.5
1
1
1
1
0.5
1
Slice interval (mm)
0.5
0.5
0.3
0.3
1
1
1
1
0.4
1
1194.4
172.9
159.9
76.2
781.8
991.1
405.5
110.9
331.8
Reconstruction protocol
Radiation dose Dose-length product (mGy ⋅ cm) 1163.2
Note—AEC = automatic exposure control.
1292
AJR:201, December 2013
Low-contrast detectability—The low-contrast sensitivity of CT scanners is routinely assessed by subjective scoring of low-contrast objects within phantom CT images. The theoretic difference in attenuation between the low-contrast objects (disks) and background in the phantom was 10, 5, and 3 HU for the 1%, 0.5%, and 0.3% contrast series. Therefore, noise, resolution, and contrast were significant factors in these measurements. Rating of low-contrast visibility by human observers can be biased because low-contrast objects are arranged in known fixed patterns. As an alternative to analysis by human observers, a nonprewhitening matched filter with eye filter model observer was applied in a two-alternative forced-choice experiment [28]. The nonprewhitening matched filter with eye filter is a mathematic model that has been found similar to human observers for detection tasks in the presence of low-pass noise [29]. Its strategy consists of correlating the image with the shape of the expected signal profile (of background and signal) filtered by the visual-response function. In a detection task, the model reaches a decision by comparing test statistics T1 (test for background) and T2 (test for signal). The test statistics are ob1.2
tained by cross correlation between the expected signal and the image [30]. An average slice was created with all the slices available for each individual dataset. Template locations (defined according to manufacturer specifications) were shifted in a 3 × 3 pixel region centered at the initial template location to find the maximum value (average pixel value within the template). The location that gave the maximum value was assigned as the new (optimized) template location. The amplitude of the signals in the template was fixed for each contrast series and based on specifications. The signal is blurred by the point spread function measured with the appropriate phantom. The eye filter, E, used in the model was E(f) = fe–bf with b chosen such that E(f) peaked at four cycles per degree. The eye filter was radially symmetric, f is spatial frequency, and e is Euler’s number. In the experiments, a fixed viewing distance of 500 mm from the monitor was assumed. From the distribution of test statistics, one can compute a discrimination index. Discrimination indexes (d′) resulting from the automated analysis were transformed into proportion correct responses varying between 0.5 (chance) and 1.0 (100% detectability)
FBP FC13
Normalized NPS
Normalized NPS
Subjective Image Quality Evaluation Subjective image quality was evaluated for 10 retrospectively selected clinical datasets. All patients
FBP FC13
AIDR 3D FC14
1.0
0.8 0.6 0.4
0.8 0.6 0.4 0.2
0.2 0.0
in a two-alternative forced-choice experiment. Objects were considered visible when the proportion correct was 75% or greater. By use of a psychometric fit to the proportion correct data points, a continuous function was obtained describing proportion correct (0.5–1.0) versus diameter (zero to infinity). The psychometric fits were performed by applying a Boltzmann fit. In our study we used 41 uncorrelated slices to obtain the discrimination index for each object. This was done 20 times for each individual datum based on a bootstrap technique. In this way the SD of just-visible object diameters could be determined as a measure of accuracy for each individual dataset. Bootstrapping is a method for assigning measures of accuracy to sample estimates. It was implemented as follows: for each individual dataset of 41 slices, 20 new datasets were constructed (equal size to the original dataset), each of which was obtained by random sampling with replacement from the original dataset (some slices present more than once in a new dataset).
1.2
AIDR 3D FC13
1.0
0
2
4
6 8 10 Resolution (line pairs cm)
12
0.0
14
0
2
4
6 8 10 Resolution (line pairs cm)
12
14
A
B
Fig. 1—Graphs show noise power spectrum (NPS) of filtered back projection (FBP) and adaptive iterative dose reduction 3D (AIDR 3D) at tube charge 75 mAs at same reconstruction kernel (FC13) (A) and at different reconstruction kernels (FC13 and FC14) (B).
0.018
0.018
FBP FC13 FBP FC14
0.016
SD of Attenuation–2 (HU)
SD of Attenuation–2 (HU)
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 06/11/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Adaptive Iterative Dose Reduction
0.014 0.012 0.010 0.008 0.006 0.004 0.002 0
0
50
100 150 200 Tube Charge (mAs)
250
300
AIDR 3D FC13 AIDR 3D FC14
0.016 0.014 0.012 0.010 0.008 0.006 0.004 0.002 0
0
50
100 200 150 Tube Charge (mAs)
250
300
A
B
Fig. 2—Graphs show inverse square of SD versus tube charge for filtered back projection (FBP) (A) and adaptive iterative dose reduction 3D (AIDR 3D) (B) at different reconstruction kernels (FC13 and FC14).
AJR:201, December 2013 1293
Joemai et al. AIDR3D
Teflon Kernel FC13
0.8
0.8 f50 (lp/mm)
f50 (lp/mm)
FBP
1.0
0.6 0.4 0.2 0.0
0
50
100 150 200 250 Tube Charge (mAs)
0.4
FBP
0.0
300 AIDR3D
0.8
0.8
0.6 0.4
0
50
100 150 200 250 Tube Charge (mAs)
Teflon Kernel FC14
1.0
f50 (lp/mm)
f50 (lp/mm)
lished by the referring physician was used. Thus the image quality score was determined for the assessment of specific clinical indications (e.g., possible metastasis, bleeding, abscess, coronary stenosis, and fractures). According to the indication, a diagnostic quality score was determined by the observers. The observers were notified that artifacts caused by motion, such as respiratory motion or cardiac motion, were excluded from the scoring. The following technical artifacts were assessed: helical or windmill artifacts, streak artifacts, and beam-hardening artifacts. Observers were instructed to express every perceived difference in image quality between FBP and AIDR 3D in five grades to record the results on a score chart. Adjustments in window settings and zooming were allowed in image evaluation. For evaluation of the subjective image quality of the clinical scans for FBP and AIDR 3D, for each CT scan the mean observer score was calculated. These mean observer scores were compared on a pairwise basis by Wilcoxon signed rank test. Variations in statistical analysis due to the different observers and clinical datasets were minimized by use of the mean observer score.
AIDR3D
0.2
LDPE Kernel FC14
FBP
300
AIDR3D
0.6 0.4 0.2
0.2 0.0
FBP
0.6
1.0
0
50
100 150 200 250 Tube Charge (mAs)
300
0.0
0
50
100 150 200 250 Tube Charge (mAs)
300
Fig. 3—Graphs of spatial resolution expressed as spatial frequency in which modulation transfer function is 50% (f 50) for filtered back projection (FBP) and adaptive iterative dose reduction 3D (AIDR 3D) show results for two contrast levels (low-density polyethylene [LDPE] and polytetrafluoroethylene [Teflon, DuPont]) and two reconstruction kernels (FC13 and FC14).
had undergone a CT examination for clinical indications according to the clinical acquisition protocol. Datasets were retrospectively selected from the most recent examinations only on the basis of the scanned region. Two studies were consecutively selected for each region: head, chest, cardiac, abdomen, and extremities. Our institutional review board does not require approval for retrospective analysis of data with patient-identifying information omitted, as was the case in this study. Five radiologists experienced in clinical CT evaluated the images. Image evaluation was performed at a dedicated medical image postprocessing workstation (Vitrea fX version 6.1, Vital Images). For
8 7
each clinical dataset, conventional FBP and AIDR 3D volumes were loaded in the workstation and presented simultaneously. The observers were blinded to reconstruction technique. The datasets were displayed next to each other at random. The radiologists assessed image quality by scrolling through the transaxial slices of two volumes (FBP and AIDR 3D) simultaneously. They graded image quality using a qualitative five-point ranking of diagnostic quality (1, excellent; 2, good; 3, moderate; 4, limited; and 5, nondiagnostic) and artifacts (1, none; 2, minor; 3, moderate; 4, severe; and 5, unacceptable). To obtain a score of diagnostic image quality, the actual indication for each CT examination estab-
FBP FC13 AIDR3D FC13
Just Visible Diameter (mm)
Just Visible Diameter (mm)
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 06/11/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
LDPE Kernel FC13
1.0
6 5 4 3 2 1 0
250
150
75 40 Tube Charge (mAs)
20
10
8 7
Results Noise NPS graphs were normalized to the maximum. These graphs showed that at equal reconstruction kernels, a small change in noise can be expected (Fig. 1A). However, NPS measurements also showed that differences in NPS shape can be reduced by use of a slightly sharper reconstruction kernel for AIDR 3D (Fig. 1B). Subjectively, noise appeared with a lower magnitude with AIDR 3D than with FBP. With both reconstruction techniques a slight increase in noise was seen when a sharper reconstruction kernel was used. It is expected that the dose is proportional to SD −2 [31]. This linear relation was tested
FBP FC14 AIDR3D FC14
6 5 4 3 2 1 0
250
150
75 40 Tube Charge (mAs)
20
A
10
B
Fig. 4—Diameters of just-visible low-contrast object for six dose levels with filtered back projection (FBP) and adaptive iterative dose reduction 3D (AIDR 3D). A, Graph shows results with reconstruction kernel FC13. B, Graph shows results for reconstruction kernel FC14.
1294
AJR:201, December 2013
for FBP and AIDR 3D (Fig. 2). FBP had the expected relation for both reconstruction kernels. However, AIDR 3D exhibited a different behavior. Data at tube charges (i.e., tube current–time product) of 75–10 mAs showed a relatively smaller increase in noise at each step in lower-dose images compared with images obtained at high tube charges (250– 75 mAs). At all dose levels, lower noise levels were found for the slightly smoother kernel (FC13) than for the sharper FC14 kernel. Spatial Resolution Spatial resolution in which MTF is 50% is shown for two contrast levels, two reconstruction techniques, and two reconstruction kernels in Figure 3. Quantitative analysis of MTF measurements showed similar performance for AIDR 3D compared with FBP. At low dose levels (< 50 mAs), however, a small decrease in spatial resolution was found with AIDR 3D. The sharper kernel (FC14) showed a slightly better resolution than the FC13 kernel. The spatial resolution was not dependent on contrast level for AIDR 3D or for FBP. Low-Contrast Detectability Figure 4 shows the diameter of the just-visible low-contrast objects (alpha), as a function of the dose level for two reconstruction kernels. As could be expected, the LCD improved rapidly at higher dose levels, ranging for AIDR 3D from detectability of objects as
Fig. 5—Graph shows image quality score averaged for individual observers. Diagnostic quality and artifacts were rated on 5-point scale (lower is better). AIDR 3D = adaptive iterative dose reduction 3D, FBP = filtered back projection.
Mean Score (Lower Is Better)
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 06/11/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Adaptive Iterative Dose Reduction 4.0 3.5 3.0 2.5 2.0 1.5 1.0
FBP AIDR 3D Diagnostic Quality
Observer 1
Observer 2
small as 1.37 mm in diameter at the highest dose level (250 mAs, FC13) to slightly greater than 4 mm in diameter at the lowest level (10 mAs, FC13). At all dose levels there was consistent improvement in the detectability of low-contrast disks with AIDR 3D compared with FBP. This means that with AIDR 3D smaller low-contrast objects could be detected at the same dose level. LCD for the FC13 reconstruction kernel was slightly better because of a lower noise level compared with the FC14 reconstruction kernel. The relative differences between FBP and AIDR 3D differed for each dose level. The best relative improvement with AIDR 3D was found at a tube charge of 10 mAs, at which 32%-smaller-diameter low-contrast disks could be detected (Fig. 4A and 4B). The best absolute improvements were found at the lowest dose
Fig. 6—Clinical examples of data used in subjective evaluation. Left, filtered back projection; right, adaptive iterative dose reduction 3D. Top, abdomen of 47-year-old man with suspected metastasis (window width, 500 HU; level, 100 HU). Bottom, shoulders of 27-year-old woman with suspected shoulder fracture (window width, 1500 HU; level, 200 HU). Slice thickness is 0.5 mm. Substantial reduction in noise and streak artifacts is evident with AIDR 3D.
Observer 3
FBP
AIDR 3D Artifacts
Observer 4
Observer 5
levels for the FC14 reconstruction kernel, with which 2.1-mm-smaller disks could be detected with AIDR 3D (Fig. 4B). Subjective Image Quality The mean scores per reconstruction technique are shown for each observer in Figure 5. The evaluation showed the same trend for all observers with an observer-dependent bias. Individual observer results showed the best diagnostic image quality and fewest artifacts for AIDR 3D. The average score across all cases and observers for diagnostic quality improved significantly from 2.4 ± 0.6 for FBP to 1.5 ± 0.3 for AIDR 3D (p < 0.0003). The amount of artifacts also decreases significantly with AIDR 3D (2.9 ± 0.7 for FBP, 1.7 ± 0.4 for AIDR 3D; p < 0.0003). An example of the performance of AIDR 3D compared with FBP is depicted in Figure 6, which shows that a marked decrease in artifacts can be realized with AIDR 3D. Discussion Objective and subjective evaluations showed consistently that image quality improves with AIDR 3D compared with FBP. Measurements revealed that LCD improves a maximum of 32% with a substantial reduction of noise with AIDR 3D compared with FBP. Subjective grading of clinical images of patients confirmed improved image quality with reduction of artifacts and potential for dose reduction of AIDR 3D. Further clinical studies should be performed to evaluate the potential dose reduction for each particular clinical indication. There is a need for methods of accurate assessment of the diagnostic performance of new imaging technologies, particularly in CT. Our approach to the assessment of spatial resolution (MTF), noise (NPS), and LCD with a human observer model is in line with the findings of subjective grading of clinical images by radiologists. Further research and develop-
AJR:201, December 2013 1295
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 06/11/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Joemai et al. ment of such methods could be of relevance for the assessment of coming advances in imaging technologies. The human observer model seems particularly promising for achieving this objective. In addition to noise measurements as SD, it is also essential to analyze the magnitude of the noise as a function of frequency on CT images. SD measurements are commonly used but have the limitation that they do not provide information on the noise appearance. As an alternative, NPS measurements are a good tool for this task. A change in the shape of the NPS indicates that the noise appearance is different. In this study the change in the NPS shape was modest, but there was substantial reduction in the magnitude of the noise. Software for automated assessment of LCD of images of a phantom was used in this study. A previous validation study [28] showed that the software model observer has high agreement with human observers. The software, as an alternative to human observers, supports objective analysis of the LCD. Figure 4 shows that diameters corresponding to the just-visible object are larger for FBP compared with AIDR 3D for 1% contrast. From these diameters the dose reduction potential of AIRD 3D can be derived while LCD is maintained, because by interpolation, the dose level that would have been required with AIDR 3D to achieve the same alpha as for FBP can be calculated [28]. For acquisitions between 20 and 250 mAs, the calculated potential dose reduction was on average 36% (range, 9–86%). Spatial resolution measurements showed no deterioration at moderate and high dose levels, but there was improvement in noise. To achieve improvement for both spatial resolution and noise, a sharper kernel can be used in combination with AIDR 3D. In this study we used a smooth (FC13) and a slightly sharper kernel (FC14). Improvement in overall image quality is a general finding in studies comparing FBP with iterative reconstruction methods, but one of the major effects of AIDR 3D was the reduction of streak artifacts on clinical CT scans, which was achieved by processing in the raw data domain. Our study had several limitations with regard to the observer study. First, the observer study was based on subjective reading, because for practical reasons a lesion detection study, which would have been more accurate, could not be performed. Second, the subjective evaluation of clinical CT images was performed side by side for the FBP and AIDR 3D images. This method introduced the dis-
1296
advantage that observers might have been able to identify the AIDR 3D image, which might have created bias. However, the results in the subjective grading were in agreement with the objective results. Side-by-side comparison was performed because our main interest was to study the difference between the two reconstruction methods, which was considered easier to achieve in direct comparison of images. The amount of dose reduction should be investigated for each clinical indication in studies with larger numbers of patients. Finally, CT examinations of the brain and abdomen are often performed with a slice thickness of 5 mm, whereas in this study 0.5- or 1.0-mm slices were used. Because this study evaluated differences between two reconstruction techniques, thin slices were used to enhance differences between the two techniques considered. Conclusion This study showed that AIDR 3D images, compared with FBP images, have lower noise, better LCD, significantly better diagnostic image quality, and significantly fewer artifacts at a similar spatial resolution and allow a potential dose reduction of 36% (range, 9–86%) to maintain the same LCD. Acknowledgments We thank A. J. van der Molen, E. L. van Persijn-van Meerten, A. Navas Canete, and M. C. Kruit for contributions to the subjective evaluation of clinical images. References 1. Deak PD, Langner O, Lell M, et al. Effects of adaptive section collimation on patient radiation dose in multisection spiral CT. Radiology 2009; 252:140–147 2. Noo F, Hoppe S, Dennerlein F, et al. A new scheme for view-dependent data differentiation in fan-beam and cone-beam computed tomography. Phys Med Biol 2007; 52:5393–5414 3. Bittencourt MS, Schmidt B, Seltmann M, et al. Iterative reconstruction in image space (IRIS) in cardiac computed tomography: initial experience. Int J Cardiovasc Imaging 2011; 27:1081–1087 4. Flicek KT, Hara AK, Silva AC, et al. Reducing the radiation dose for CT colonography using adaptive statistical iterative reconstruction: a pilot study. AJR 2010; 195:126–131 5. Gosling O, Loader R, Venables P, et al. A comparison of radiation doses between state-of-the-art multislice CT coronary angiography with iterative reconstruction, multislice CT coronary angiography with standard filtered back-projection and invasive diagnostic coronary angiography. Heart 2010; 96:922–926
6. Leipsic J, Labounty TM, Heilbron B, et al. Adaptive statistical iterative reconstruction: assessment of image noise and image quality in coronary CT angiography. AJR 2010; 195:649–654 7. Cornfeld D, Israel G, Detroy E, et al. Impact of adaptive statistical iterative reconstruction (ASIR) on radiation dose and image quality in aortic dissection studies: a qualitative and quantitative analysis. AJR 2011; 196:[web]W336–W340 8. Marin D, Nelson RC, Schindera ST, et al. Lowtube-voltage, high-tube-current multidetector abdominal CT: improved image quality and decreased radiation dose with adaptive statistical iterative reconstruction algorithm–initial clinical experience. Radiology 2010; 254:145–153 9. Singh S, Kalra MK, Hsieh J, et al. Abdominal CT: comparison of adaptive statistical iterative and filtered back projection reconstruction techniques. Radiology 2010; 257:373–383 10. Schindera ST, Diedrichsen L, Muller HC, et al. Iterative reconstruction algorithm for abdominal multidetector CT at different tube voltages: assessment of diagnostic accuracy, image quality, and radiation dose in a phantom study. Radiology 2011; 260:454–462 11. Winklehner A, Karlo C, Puippe G, et al. Raw data-based iterative reconstruction in body CTA: evaluation of radiation dose saving potential. Eur Radiol 2011; 21:2521–2526 12. Moscariello A, Takx RA, Schoepf UJ, et al. Coronary CT angiography: image quality, diagnostic accuracy, and potential for radiation dose reduction using a novel iterative image reconstruction technique—comparison with traditional filtered back projection. Eur Radiol 2011; 21:2130–2138 13. Funama Y, Taguchi K, Utsunomiya D, et al. Combination of a low-tube-voltage technique with hybrid iterative reconstruction (iDose) algorithm at coronary computed tomographic angiography. J Comput Assist Tomogr 2011; 35:480–485 14. Habets J, Symersky P, de Mol BA, Mali WP, Leiner T, Budde RP. A novel iterative reconstruction algorithm allows reduced dose multidetector-row CT imaging of mechanical prosthetic heart valves. Int J Cardiovasc Imaging 2012; 28:1567–1575 15. Gervaise A, Osemont B, Lecocq S, et al. CT image quality improvement using adaptive iterative dose reduction with wide-volume acquisition on 320-detector CT. Eur Radiol 2012; 22:295–301 16. Yamada Y, Jinzaki M, Hosokawa T, et al. Dose reduction in chest CT: comparison of the adaptive iterative dose reduction 3D, adaptive iterative dose reduction, and filtered back projection reconstruction techniques. Eur J Radiol 2012; 81:4185–4195 17. Yoo RE, Park EA, Lee W, et al. Image quality of adaptive iterative dose reduction 3D of coronary CT angiography of 640-slice CT: comparison with filtered back-projection. Int J Cardiovasc Im-
AJR:201, December 2013
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 06/11/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Adaptive Iterative Dose Reduction aging 2013; 29:669–676 18. Ohno Y, Takenaka D, Kanda T, et al. Adaptive iterative dose reduction using 3D processing for reduced- and low-dose pulmonary CT: comparison with standard-dose CT for image noise reduction and radiological findings. AJR 2012; 199[web]:W477–W485 19. Juri H, Matsuki M, Itou Y, et al. Initial experience with adaptive iterative dose reduction 3D to reduce radiation dose in computed tomographic urography. J Comput Assist Tomogr 2013; 37:52–57 20. Ghetti C, Ortenzia O, Serreli G. CT iterative reconstruction in image space: a phantom study. Phys Med 2012; 28:161–165 21. Sagara Y, Hara AK, Pavlicek W, et al. Abdominal CT: comparison of low-dose CT with adaptive statistical iterative reconstruction and routinedose CT with filtered back projection in 53 patients. AJR 2010; 195:713–719 22. Prakash P, Kalra MK, Ackman JB, et al. Diffuse lung disease: CT of the chest with adaptive statis-
tical iterative reconstruction technique. Radiology 2010; 256:261–269 23. Yang Z, Zamyatin AA, Akino N. Effective datadomain noise and streak reduction for x-ray CT. In: 11th international meeting on fully three-dimensional image reconstruction in radiology and nuclear medicine 2011:290–296. Fully3D website. http://www.fully3d.org/2011/Fully3D2011Proceedings.pdf. Accessed August 5, 2013 24. Yang Z, Silver MD, Noshi Y. Adaptive weighted anisotropic diffusion for computed tomography denoising. In: 11th international meeting on fully three-dimensional image reconstruction in radiology and nuclear medicine 2011:210–214. Fully3D website. http://www.fully3d.org/2011/Fully3D2011Proceedings.pdf. Accessed August 5, 2013 25. Boedeker KL, Cooper VN, McNitt-Gray MF. Application of the noise power spectrum in modern diagnostic MDCT. Part 1. Measurement of noise power spectra and noise equivalent quanta. Phys Med Biol 2007; 52:4027–4046
26. Joemai RM, Geleijns J, Veldkamp WJ. Development and validation of a low dose simulator for computed tomography. Eur Radiol 2010; 20:958–966 27. Richard S, Husarik DB, Yadava G, et al. Towards task-based assessment of CT performance: system and object MTF across different reconstruction algorithms. Med Phys 2012; 39:4115–4122 28. Hernandez-Giron I, Geleijns J, Calzado A, et al. Automated assessment of low contrast sensitivity for CT systems using a model observer. Med Phys 2011; 38:S25–S35 29. Beutel J, Kundel HL, Van Metter RL. Handbook of medical imaging. Volume 1. Physics and psychophysics. Bellingham, WA: SPIE, 2000 30. Reiser I, Nishikawa RM. Identification of simulated microcalcifications in white noise and mammographic backgrounds. Med Phys 2006; 33:2905–2911 31. Primak AN, McCollough CH, Bruesewitz MR, et al. Relationship between noise, dose, and pitch in cardiac multi-detector row CT. RadioGraphics 2006; 26:1785–1794
AJR:201, December 2013 1297
